Google’s new TurboQuant compression for large language models has rattled markets, but the fallout is more complicated than a quick sell-off. While memory-chip names dipped on the news, TurboQuant could actually accelerate demand by making it practical to build much larger models. This article walks through how a compression breakthrough can spur a memory supercycle and why Micron, Sandisk, and Seagate deserve a second look.
TurboQuant is a clever way to shrink the memory footprint of huge language models without sacrificing accuracy, and Google laid out the idea in a recent technical post. On the surface that sounds like a threat to memory makers, since less memory per model might imply lower chip demand. The key mistake is treating memory consumption per model as a fixed ceiling instead of a cost variable that reshapes behavior.
When the market first digested the TurboQuant write-up, investors punished several memory names hard and fast. Shares of Micron Technology, Sandisk, and Seagate were hit as traders rushed to price in a potential drop in revenue tied to memory shortages easing. That knee-jerk reaction ignored the bigger dynamic brewing in AI infrastructure and data-center economics.
Here’s the counterintuitive part: making memory cheaper or denser often leads teams to build bigger, more capable models rather than simply running the old ones cheaper. Model size has exploded; the largest LLM listed in 2019 had about 0.09 billion parameters, which ballooned to roughly 540 billion by 2022. Today many leading models push into the hundreds of billions or over a trillion parameters, and removing a memory constraint unlocks even larger architectures.
Lower effective memory costs change the optimization calculus for AI labs. If training or inference runs drop substantially in price, developers will use that margin to expand model depth, add modalities, or increase dataset size to chase accuracy and capabilities. “Will AI create the world’s first trillionaire?” is a provocative line the industry is asking because the economic upside of larger, smarter models can be enormous and drive demand for more storage and memory capacity, not less.
Industry research also points to sharp cost declines for inference over time, driven by better chips and utilization. That kind of cost compression makes previously expensive experiments routine, and routine experiments generate much more data to store and process. The practical outcome is a rising tide of data-center workloads that require DRAM, NAND, and high-capacity HDDs—precisely the products Micron, Sandisk, and Seagate sell.
Each of the three companies sits in a different niche of the memory stack: Micron covers compute-centric DRAM and NAND flash, Sandisk is focused on NAND flash storage, and Seagate supplies high-capacity hard-disk drives for bulk storage. Those product differences matter because AI workloads use a mix of ultra-fast memory for compute and cost-efficient, high-density storage for datasets and archives, creating multiple demand vectors across the market.
Data growth from AI is staggering on paper and in practice. One projection shows AI-fueled applications generating 394 zettabytes of data in 2028 versus 72 zettabytes in 2020, an implied compound annual growth rate near 24 percent. Even if these are directional estimates, they illustrate the scale of the opportunity: more models, more training runs, more checkpoints, and more backups all translate into sustained demand for chips and drives.
Valuation math matters too. Today Seagate trades near 24 times forward earnings—slightly above the tech-heavy Nasdaq-100’s forward multiple—while analysts expect sizable earnings gains in the next two fiscal years. Micron’s forward multiple sits much lower around the mid-single digits, and Sandisk appears closer to a mid-teens multiple, making both firms look relatively inexpensive against potential growth. Cheap multiples combined with robust secular tailwinds create a clear thesis for investors willing to ride cycles.
That said, buying any memory stock requires respect for cyclicality and execution risk. Capacity investment timing, yield curves, and end-market demand swings can swing earnings violently. If you consider adding shares, weigh company-specific operational strength and balance-sheet flexibility alongside the bullish structural story for AI-driven memory demand.
